Seagrass detection in the mediterranean: A supervised learning approach

We deal with the problem of detecting seagrass presence/absence and distinguishing seagrass families in the
Mediterranean via supervised learning methods. By merging datasets about seagrass presence and other external
environmental variables, we develop suitable training data, enhanced by seagrass absence data algorithmically
produced based on certain hypotheses. Experiments comparing several popular classification algorithms yield up
to 93.4% accuracy in detecting seagrass presence. In a feature strength analysis, the most important variables
determining presence–absence are found to be Chlorophyll-α levels and Distance-to-Coast. For determining family,
variables cannot be easily singled out; several different variables seem to be of importance, with
Chlorophyll-α surpassing all others. In both problems, tree-based classification algorithms perform better than
others, with Random Forest being the most effective. Hidden preferences reveal that Cymodocea and Posidonia
favor the low, limited-range chlorophyll-α levels (< 0.5 mg/m³), Halophila tolerates higher salinities (> 39),
while Ruppia prefers euryhaline conditions (37.5–39).